💬｜general-chat

[stone latch]

I wish 😂

[pearl ocean]

GTA6

[foggy halo]

real 🥲

wish lol 😂

[pearl ocean]

I use a 4080

[shell tendon]

16gb is muuuuch better than 12

[karmic cedar]

quality-wise 16gb is a make-or-break window for a lot of folks

[shell tendon]

yeah

you're slamming into OOM far less often

[astral goblet]

if you're buying a gpu today, anything less than 16gb is a joke

[shell tendon]

you're def crazy to go under 12gb

better off skipping the daily to go coffee for a month or two or whatever and at least getting a 4070 super

[trim magnet]

if u using a lot of ai related stuff less than 48gb is a joke

[rich kestrel]

buying a gpu today period is a joke, when 5090 is coming out

[shell tendon]

that might not come out for a year

[rich kestrel]

wrong. coming out end of year

[shell tendon]

you don't know that, lol

[trim magnet]

if u dont wanna wait u can always buy an A6000

[shell tendon]

i bought a 4090 a month or two ago knowing the 5090 would be out in about a year, maybe a lil longer, cuz life is short and it gives me an excuse to buy another killer card and set up a second system when it comes out

[astral goblet]

5090 is still going to be 24gb. the TMSC forges aren't cranking out denser chips and nvidia isn't going ot eat their enterprise markets

supply and demand

5090 is also technically still a gaming card. Games, even at 4k, won't need more

yeah also what clownshark said. if yo'ure not poor then it's something to look forward to

[trim magnet]

pretty sure if u are talking about being poor then a 5090 is not something to flex about

[astral goblet]

sounds like something a poor person would say

/breaths the copium so deep/

[trim magnet]

you right a poor person would flex about a 4090 on a random chat online

[astral goblet]

if they spent next months rent on it and need to convince themselves it was worth it, i bet they would

[trim magnet]

maybe they bought one of those custom versions that are really expensive that have like gold plated fans

theres ram like that too,with some gold on heatsink

[shell tendon]

yeah, my concern is they'll use it as an excuse to leave it at 24gb of vram

if it's still 24gb i won't buy it unless it's at least 50% faster

if it's 32gb or more, i'm buying it

[astral goblet]

it'll be a lot faster for sure. i don't think it'll have higher vram. i'm confident in this prediction but i'd be happy to be wrong

[shell tendon]

yeah, no idea what to expect there

i've heard a lot of contradictory shit and it doesn't make a diff if i spend all day reading about it trying to sort that out, so i don't bother

it'll come out when it comes out, till then, i'm enjoying the absolute fuck out of my 4090

if it's got more vram, great, if not, well then whatever, f nvidia and hopefully the competition finally catches up and kicks their asses like amd did to intel when they got fat and lazy

[astral goblet]

i might trade my 4080 up. i only got the 4080 because prices of gpus were insane a year ago and the 4080 was the only one canadian retailers weren't pumping the price on. it was still msrp.

[shell tendon]

ahh yeah makes sense

[astral goblet]

i'm only so not cheap

[shell tendon]

i managed to score my 4090 at msrp which was really lucky

[astral goblet]

i'm still pretty cheap

[shell tendon]

yeah, i'm not ever going to pay a fn scalper

[astral goblet]

absolutely

[shell tendon]

but yea the way i see it is if my 4090 is worth literally nothing in 2 years, that's 3 dollars a day and i'm getting a hell of a lot of entertainment for 3 dollars a day

plus, AI is the future for sure and i'm at least getting to tinker with some stuff related to it and get familiar with some important technology that i'd otherwise probbaly be too tired to bother with after work

3 dollars a day = what i would prolly spend on a nice IPA watching a movie in front of the tV instead anyway... really not much

[astral goblet]

i used to lower my field of view to get more consistent frame rates. my 4080 though, i kept that habit up and it was blowing smooth framerates at me so hard that the narrow field of view was causing me serious motion sickness. can't say i've ever had that gpu problem before

[shell tendon]

yuck

motion sickness

so now you've got that expanded i'm guessing

[astral goblet]

like i'd had to just walk away from the pc immediately, get fresh air, burp a ton, it sucked. i'm like "Wtf mate?"

[shell tendon]

framerates are a weird one with that

[astral goblet]

yeah i turn up the fov and it doesn't happen no more

if i'm not careful when in helicoptor situations though. 🤢

got that on halo reach recently

[shell tendon]

the only game that has ever given me motion sickness was the new ratchet and clank

it was only for a moment, but it did actually happen

i'm lucky for that

not that i'm immune... boats... oh god

[astral goblet]

when i first started using vr , framerates were so bad that i got it hard

[shell tendon]

ahh i haven't used vr

[astral goblet]

yeah i'll get sea sickness hard and i live on an island

[shell tendon]

ugh

last time i went out in real waves, it took me over two weeks to feel right

haven't gone back out and prolly won't ever again lol

[astral goblet]

i can't do carnival rides. not the intense ones

[shell tendon]

yeah i'll pass on that too

[astral goblet]

tea cups MAYBE

[shell tendon]

there was a six flags incident when i was a kid and i haven't had much interest in any of that siht since

my cousin went on the batman ride

where you're strapped in by the shoulders and it goes in loops, etc, that's all that's holding you in

[astral goblet]

i sat at the front of the "sky master" and haven't done any rides since

[shell tendon]

and the damn thing DIED and they got stuck upside down at the top of a loop for i think over an hour

ppl were throwing up and it was raining down everywhere

i was done lol

[astral goblet]

https://youtu.be/_mOGVOVZw70 this guy ended my carnival days

[shell tendon]

ugh that looks baaad

[astral goblet]

yeah. lol. so while i'll deal with motion sickness in a lot of situations, i've never had it in pc games till my 120hz screen was fed with all the frames

fun times

been gaming since 86

[shell tendon]

damn, yeah

hilarious thing to me is i have yet to fire up a game on this 4090

i couldn't have imagined a few months ago that i'd get a 4090 and not play any games on it

[astral goblet]

try out "the finals" it has wicked crazy destructable environments, runs on unreal engine 5, and is free

[shell tendon]

i may have to try that sometime

backlog is a disaster

[astral goblet]

throw it all out. none of it is unreal engine 5

[shell tendon]

at least my gamepass sub (got ultimate for 5/mo 3 years ago) finally expired. that solves one problem

[astral goblet]

star citizen is almost here i heard. another 2 years maybe

[shell tendon]

alan wake 2 is on my list but i still need to finish 1

loved control

about halfway through BG3

2/3 through horizon forbidden west

[astral goblet]

control i started playing. was fun. then i got distracted and forgot

[shell tendon]

halfway through the last of the mass effect trilogy

[astral goblet]

is probably one of the games eating 70% of my drive space

[shell tendon]

control is as cool as it gets

damn i love that one

it's the atomsphere and the humor that gets me

the sheer weirdness of it

[astral goblet]

red dead redemption 2 is a worth while title. i've sunk most time inot that lately. it's so pretty and runs so well on the ada cards

[shell tendon]

prolly not surprising if you've seen how weird my art on here is

oh god, that's another one i got halfway through

[astral goblet]

oh and oxygen not included. not really a graphic flex.

[shell tendon]

generally the issue isn't burning out on a game, it's gcetting a work crunch for a month or two, then by the time that passes i fire up something new because i'm no longer on a roll with the last game

[astral goblet]

hear that

"got distracted" is my way of dancing around crippling work hours

no mans sky is really pretty too. got a huge update AGAIN just recently

[karmic cedar]

but can you prepare duck confit in no man’s sky?

that takes a certain level of culinary talent.

[shell tendon]

yup, the problem is i come home so exhausted i don't even have the energy to game

sometimes i'll try, but i can't focus enough to figure out what i'm doing in a strategy or open world game, nad lack the reflexes for anything else

[astral goblet]

focus? i hardly know us

[shell tendon]

ha

[karmic cedar]

fooocus

[low moon]

foooooooooooocus

with each update they could ad an o

[karmic cedar]

missed oppooooooooortunity

[astral goblet]

"are you using fooocus?" "no i'm on foooocus"

[low moon]

Still? But foooooocus is out!

Imagine one day you come in and in "Stable Models" you see the SD3 room XD

[astral goblet]

i still have doubts that the board will be allowed to release the model. they can make more money by selling it and the investors might demand that

emads all about blockchains now too so you know

you know how blockchains go

[shell tendon]

yeah, i'm a lil concerned about that too

[pearl ocean]

any nice Civitai Models to get??

[astral goblet]

all the existing tools are still here. can't unrelease weights or code really

[pearl ocean]

we don't talk about dreams on this server

[astral goblet]

sd 15 "cyber realistic" has new versions out. pretty nice but still very sd15

[low moon]

Is it just me or SD15. while not bad feels more like video and sdxl feels more mature and film like

[astral goblet]

i've found proteus to be VERY high quality generations. supposedly it's based on ponyxl.

[pearl ocean]

im think im using CyberRealistic v42

is there a newer one?

[astral goblet]

thats the one

[low moon]

if they dont release sd3 probably the community will refine cascade somehow

[pearl ocean]

its the MOST realistic model I think

[low moon]

cyberrealistic is the best? not photogasm?

[pearl ocean]

photogasm comes in 2nd I think

then realisticvision

[astral goblet]

there's a 2.5d cyberrealistic too but i didn't think it worked too well

[brittle geyser]

Okay is it just me or not using a dozen random prompts like high quality, low quality, and all that jazz actually result with a much better photo generation than trying to shoehorn all that stuff in?

You some of my best generations didn't have anything like that in it.

[low moon]

its a vague nebulous world, no definitive answers, the magic and bane of ai image generation, its still very much a random slot machine spin

occasionally it blows your mind but even then not much you can do with it as there is no real consistency yet

[astral goblet]

depends on your use case. if yo'ure using a model thats really refined to the content you prefer, you might find prompting to be easier

turns out a lot of people have very similar use cases and there are a lot of models refined towards their "needs"

[brittle geyser]

I'm not so sure about that. Because I found if you go into a model and try and force it into a image style that is not trained on, it requires a lot of those pointless keywords. But if you find the model that you like and just allow it to do its thing it does great.

Yes it is a bit random but I find it using less keywords but more along the lines of the proper keywords is where you can actually control it fairly well.

At least in my personal experience so far.

I found that a lot of those keywords kind of... I guess you could say make everything similar. And it also takes away a lot of the magic of the randomization that your other keywords give you. It kind of gives you a narrow range for the AI to work with if that makes any sense?

[astral goblet]

i use prompt magic in the dynamic prompts extension to add all that flavor text to prompts using a gpt2 model

[brittle geyser]

Prompt magic is fun.

[low moon]

i can say for sure SDXL needs less rpomting wisardry

[brittle geyser]

Definitely can throw some variety in there. I've really enjoyed started using artist or photography keywords in my promp ts and those end up really fun.

[astral goblet]

artist names are always good flavor

[low moon]

pointillist obese cat floating in space

[brittle geyser]

I guess what I'm getting at though is like using a set of words like high resolution high quality masterpiece the forces it to use source images based on that where I don't think every single beneficial image would have those keywords so it excludes potential benefits.

I feel like those keywords are an attempt to create a perfect image on the first generation and not subsequent image to image or edits.

I feel like the first few generations of an image are really just to get the basics compositional down. Try not to force it to do too much in one pass. Do it on iterative generations bit by bit

[low moon]

its hard to spend hour son perfecting an image even tho if done like that probably any image could be created that you cna imagine but the temptation to keep trying with new promts an the ease at whcih ti can be done is too much

[brittle geyser]

Well really not so much a perfect image in the first pass but it feels like that they're trying to get to as close as perfect on the first pass as possible.

And yes I do agree. But it's just getting kind of silly in some of the source images that I've seen. Like they have like 150 tokens worth of negative prompt keywords and I think that's just getting a bit silly.

[low moon]

well foooocus ads loads of tokens in those styles

like hundreds

[brittle geyser]

Focus?

I've been using a lot of forge recently

[astral goblet]

wait until people have t5 context length. then you'll see some crazy prompt metas

i'm gonna drop zepellin lyrics in every image

[brittle geyser]

Geez I can only imagine. Now I just want to copy and paste my favorite song lyrics into an image and just see what comes out. Wish I was home at my computer. Out of town at the moment

Definitely going to do some tool lyrics.

You've given me a new idea my friend.

[astral goblet]

if any of the ui's had lavi bridge adapters implemented, i'd recommend using t5 on your fave sd15 model

but what are you gonna do?

[brittle geyser]

What's t5 and what's a lavi bridge adapter?

[wanton condor]

is the video API down for anyone?

[astral goblet]

t5 is a text encoder model that will be used in sd3

[brittle geyser]

Oh I'm just a casual man. I'm technically inclined and everything but I'm still pretty basic when it comes to the advanced stuff on stable.

I so basically what I'm saying I don't know what you mean by text encoder.

[astral goblet]

clip only has 75 context length. it can only understand that much in any context. while it can do longer prompts, it'll only understand it in 75 chunks

[brittle geyser]

Right I follow I'm familiar with that part

[astral goblet]

t5 has something like 800

[brittle geyser]

Oh interesting.

[astral goblet]

lavi bridge is one of the new models that have come out to create embeddings for sd15

[brittle geyser]

But will it be able to understand natural language better?

Because that's something that I really look forward to. The ability to"paint a picture" in words and get a accurate output

[low moon]

Just type in "cool movie" and get a full lenght feature film

[brittle geyser]

So I look forward to when it becomes less about prompts and keywords and you can effectively describe what you wish for

[astral goblet]

one of the experimental models out there uses t5 already. i forget which one.

[brittle geyser]

Yeah could be a full length feature film but that ain't going to get you a shit what you were hoping for. You got to be detailed man lol

[low moon]

exactly

henc ewe need consistency

[astral goblet]

https://github.com/PixArt-alpha/PixArt-alpha this model uses t5 but a very small training dataset. you can sort of catch a glimpse of what t5 can understand

[brittle geyser]

Right but it natural language let's say I did this. I wanted a 1967 mustang with a large supercharged engine with flames shooting out the exhaust, next to a futuristic sports car wreath an electrical discharge.

Currently it's going to get all confused. But if it understood it in a natural language sense it would be able to differentiate the separate vehicles and effects.

It's going to end up with both cars with both of the features. A natural language model would be able to understand the you want new next to old fire next to electricity.

[astral goblet]

natural language works in sdxl reasonably well if you feed it to the openclip layer instead of both

sdxl has 2 text encoder layers and sd3 has 3

[brittle geyser]

Huh ... Ok I didn't know that

I've been playing around with large language models and stable diffusion a bit but again, I'm kind of casual. I guess I'm above the general pop in understanding.. But I'm still not high level understanding

And I don't really have a super lot of experience in SDXL. I'm running on the 3080 and it's kind of hard to make the transition with the speed change

[low moon]

No worries soon, we will try to force it to do the bizarre trippy stuff it does now. When every prompt will yield very realistic normal images.

It will be called Retro AI

AIstehetic

"A subgenre trying to replicate early AI image generation artifacts"

"Deformed hands is one of its staple"

[astral goblet]

higher base resolution though. if you're doing megapixel 1.5 2 pass generations, its comparable speeds

https://www.reddit.com/r/StableDiffusion/comments/15c2n0q/sdxl_two_text_encoders_two_text_prompts/ thread on the dual layers . you can really only do them in comfyui i think. the auto1111 devs scoff at it and don't think they work.

you can get a really good idea of the natural language understanding in the openclip layer by playing around with the 768x 2.1 model. it's built on openclip and doesn't use the Vit-L model at all

the clip ViT-G model is the original pretrained model that sd15 was trained on. it was made with REALLY bad data but capable data. That's where all the flavor text capability is coming from. Poeple learned to throw special tokens at it because it worked.

But 2 years later.. sd15 models are REALLY refined with a lot better data. so while that text encoder is still pretty weak, its more refined than it was

[karmic cedar]

I really want to use OpenAI’s voice model to turn my speech into Jamaican patois

[astral goblet]

I want them to do harry belefonte's version of rap god

this emo demo has got me hungry for more rap god https://humanaigc.github.io/emote-portrait-alive/content/video/16比9视频结果/song_cxk.mp4

[karmic cedar]

This reminds me of the initiative to fine-tune a language model on Google’s product names. “DeepName”

🥁

I’m thinking about tomorrow.

[astral goblet]

my plan? right now

[karmic cedar]

excuse me

we need to talk about this

[astral goblet]

i thought you wanted to release it now wtf

[karmic cedar]

you need to slow down and consider the shareholders a little bit more like i told you on the phone

[astral goblet]

that wasn't me on the phone bro i think you got ai scammed

[karmic cedar]

bullshit that’s not possible yet

[astral goblet]

open ai trolling you with their unreleased voice thing

[karmic cedar]

well, i guess that settles things

no SD3

[astral goblet]

of course they'd want sd3 released tommorrow. they'll delay the bomb all they can

[karmic cedar]

they’re fishing with the best bait they’ve got. but OpenAI’s over there with their aged ahi tuna steaks

[astral goblet]

fuck i want some ahi tuna now

[karmic cedar]

me too. hold the mercury

[astral goblet]

damn youuu sd3!

nothing wrong with a little heavy metal 🤘

[pearl ocean]

maybe one day

[astral goblet]

roku's basiliisk is actually a diffusion-transformer model? hmm

[pearl ocean]

Imagine how realistic SD3 will look

[low moon]

Maclunky.

[wispy storm]

Wha is your guys favorites GUI for SDXL?

[loud solar]

Stable Diffusion Forge ...

[open vigil]

is not about realistic but the image resolution so the dettail and over all the most important is the way the training of models is but im not expert

[latent spade]

Will SD3 be released to the community or are they closing it up?

[honest spear]

last news I saw from the stability ceo was that they're improving security without compromising quality, then beta, api and then weights. It's taking soo long

[dim surge]

Hello, how do linux distros work with SD? is it true more vram is needed

[loud solar]

The more VRAM, the better

[dim surge]

understood. But the question is whether *nix system are less optimized

[loud solar]

I would think they should work better. But I am not really a pro in that topic

[charred mesa]

3-5 weeks ETA (4-6 weeks but this week is almost over)

[honest spear]

It's just deep floyd / IF all over again, they say soon, then they mean 3 months. I personally kill time buying and trying 3d assets, at worst I'll have a bunch of things I can throw to a canny preprocess

[charred mesa]

yeah its impossible to know, I just wrote whatever the CTO wrote

[static cape]

"improving security without compromising quality" what does that even mean... sounds like shallow Corp talk to me.

[charred mesa]

^

maybe safety actually

maybe it's the censorship of their services and API

they can't do much to the model itself

[mossy bone]

^

Lorem ipsum dolor sit amet

consectetur adipiscing elit

sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.

[opal hedge]

no satanic chants in chat please

[charred mesa]

fr fr lorem ipsum my beloved

[narrow kernel]

It means censoring the model, which it needs to be, and probably making it harder for degenerates to make NSFW models so easily. Probably means some of the training data and weights will remain closed

[trail lion]

I have faith the degenerates will be victorious

[hallow mesa]

How does the Easter Bunny stay in shape?

Eggs-ercise ! 🥚

Happy Easter folks !

[turbid bay]

Includes a lot of things that are done to avoid causing headlines kind of like Google's image generation did a while ago

Part of it is making it difficult to generate pornographic content, of course.

But also things like removing biases the model inherits through its training data and making the model more unlikely to generate offensive content on its own.

I don't know what extent they will go to. It might also include things like preventing the model from generating likenesses of public figures and preventing the infringement of IPs.

[raven agate]

I explained why removing biases wouldn't work and told everyone exactly what would happen if they tried lol. Organic biases make the model work as intended, forced biases cause issues. Google proved exactly what I stated back then to be true 😅.

[turbid bay]

It's a difficult problem for sure and can't be fixed as naively as Google attempted.

[raven agate]

Openais approach wasn't any better. The complaints about GPT 4 losing functionality and quality is because associations have been manually altered lol

[turbid bay]

It wasn't better for the users, but it was probably better for OpenAI

[raven agate]

If you take Obsidian notes, map your seed output with the seed, the image results, the prompt, and settings. You can create a map of the latest space. There's a display option to see how all the data connects. I've used this to map out SD 1.5s latent space and if you do that, you'll see why altering the biases is problematic. The way things connect for context is amazing.

[turbid bay]

The issue is that image generation models aren't being treated the same way as i.e. photo editing programs by the public. If someone makes an offensive image using Photoshop, then nobody would reasonably blame Adobe, but image generation technology is viewed differently in the public eye and therefore companies need to take some level of responsibility for the outputs their models generate.

[turbid bay]

Sounds interesting

[raven agate]

It's cool, the local opensource stuff has been outperforming again so that makes me happy lol.

[turbid bay]

100% same

[turbid bay]

Do you have an example of a map like that?

[raven agate]

I'm glad there's an article about Emad being a Robin hood type lol

[turbid bay]

StabilityAI sure is a light at the end of the tunnel for now

[raven agate]

Not off hand, but check this out. https://support.noduslabs.com/hc/en-us/articles/360015660099-How-to-Import-and-Visualize-Your-Roam-Research-Obsidian-and-Zettelkasten-Markdown-Format-Notes

Obsidian notes made my life so much easier with AI and ML research lol

[turbid bay]

Mhm, I actually use it to take lecture notes

[raven agate]

I started using it for DnD and saving prompts.

[turbid bay]

I don't really make use of the hyperlinking too much, I use it as more of a Markdown + Latex editor tbh

[raven agate]

Then I noticed all the connections like I did with Spectrograms when I figured out how to do voice with Riffusion lol.

After that, you kind of know how to draw out the data you need as long as they don't force associations, you can predict the AI lol.

But if you know all the forced associations, you can predict it still I guess.

I had to study psychology, physics, python, and bunch of other crap I didn't plan on learning just to make my art into music videos lol.

[mortal delta]

Is it possible to create more than 3 seconds of video without things looking weird? also im not sure what video extension to use for automatic1111.

[loud solar]

SVD is made for up to 25 frames ...

[mortal delta]

Ive been looking into different extension and svd seemed the most promising but i just wanted to get the communities option and such.

[loud solar]

I always forget the stupid name of the extension I have been working with ^^

[mortal delta]

sometimes i forget aswell.

[loud solar]

Deforum ...

[mortal delta]

and what does deforum do, isint that option considered outdated or am i wrong?

[loud solar]

Complete different technology cause it's rendering frame by frame ... less consistance ...

You can see it in a lot of music videos ...

[mortal delta]

so i guess it might be considered outdated, also ai video seems kinda annoying because of how much time and effort it will take to create soemthing.

[raven agate]

I stitch together SVD, Pika Lab, and RunwayML videos or use animations as a base and turn them into more coherent videos with AnimateDiff.

[loud solar]

Not really outdated ... you can use video2video to get more control ...

[mortal delta]

that seems like a ton of work...

[mortal delta]

Interesting,

[loud solar]

As it's always been 🙂

[raven agate]

https://civitai.com/models/367412/geeky-ghost-vid2vid-organized-v1

[mortal delta]

for some reason i really want to make an anime series now with ai, but it seems impossible at the moment.

[raven agate]

Nah, takes little to no time with my workflows lol

The wolf is probably the best example, I made the video in pika, used it as a base to transform the rest using ip adapters.

[mortal delta]

how do i even use this workflow and does it work with 1111?

[raven agate]

Use the comfyui extension in Automatic1111

Or use comfyUI itself. It's similar to automatic but modular and faster.

[mortal delta]

that exists, dang i have alot to learn.

[honest spear]

Stability jap said they're working with anime studios to give them some tools, if we're lucky they'll release something in the long run

[raven agate]

I have an anime workflow 😅

A grafiti comic one as well.

You can do it all right now

ComfyUI is best for it. There's a massive community with workflows and tutorials.

[mortal delta]

how many workflows do you even have?

[raven agate]

Video offline is mainly from comfyUI AnimateDiff and Deforum these days.

[raven agate]

A few hundred?

I have way more models lol

[mortal delta]

Impressive!

[opal hedge]

Impressive indeed

I only use one workflow for everything that I keep upgrading

[raven agate]

How I'd recommend doing anime right now would be... using blender, daz3D, and make low quality animated videos or use stock videos or take your phone and record you and your friends acting out the scene you want.

[karmic cedar]

don’t judge a motion model by the amount of keyframes it produces! ask not what frames it can produce, ask what motion it can conjure!

[opal hedge]

slight zoom in and out

[raven agate]

Train or download a Lora for anime

Generate images with the lora using frames from the video or animation you made as an int for image to image.

[opal hedge]

Hopefully we get a contender for Sora on SD sometime next year. Maybe even this year if we're lucky

[karmic cedar]

we = the ones without $50 media generation plans for Company X/Y/Z

[raven agate]

Use the video as an int, use the stylized image to image in the IP adapters, use the anime lora.

Take the output and use wave2lip to create the lipsync.

[mortal delta]

well i litteraly cant 3d model and i have 0 friends.

[raven agate]

Use Bark or one of the RCVs to make the voices.

[raven agate]

Don't need to, just use a base mesh.

Use Daz3D free assets 😅

[mortal delta]

i think bark might be a good choice.

[raven agate]

If you want music use SunoAI

Or MusicGen

[mortal delta]

oh yeah, also daz only supports Nvidia and i have amd.

[karmic cedar]

https://github.com/jasonppy/VoiceCraft for voices

i’m 98% certain this is what prompted OpenAI to announce Voice Engine

[raven agate]

Hmm, 3Delight should still render for you right?

Like you can't use irat

Iray

But you wouldn't need that, you can literally render the view finder for this 😅

[mortal delta]

oh...

[raven agate]

Use make human and blender

Use um, facerig

The facerig route let's you record video of your virtual avatar talking, Use that video lol.

Like everything's a tool right now thanks to AI. All the old software is valid.

[mortal delta]

this is alot of info.

[raven agate]

Wings3D will let you make epic spaceships in AI

Take game footage from you playing a game lol

[karmic cedar]

It’s easy to forget all of the raw granular power old software has

[mortal delta]

thats sneaky but cool

[raven agate]

Everything is data and all data can now be used lol

For something

Lol

[loud solar]

You could record Scenes in Secondlife and transform using A.I.

[raven agate]

Yep

[mortal delta]

ill probably use game footage and and stock video along with other stuff if i can.

[karmic cedar]

Companies like Adobe are going to move towards having an agentic UI where you’re just talking to Clippy about your photo the whole time. I think low key they’re going to try and negate the use of manual healing and brushwork tools as those are going to become harder for an AI model to detect as being artificial.

[raven agate]

Games are now studio spaces where you can act out your movies lol

[mortal delta]

dont people also use the sims?

[raven agate]

Turn off the UI, capture screen, input galore lol.

[loud solar]

I don't know ... only have Secondlife 🙂

[mortal delta]

oh ok

[raven agate]

Mine craft can be input for crying out loud lol

[mortal delta]

i wonder what the og doom game would do with stable diffusion?

[raven agate]

I was tempted to apply for the palworld comic position

[karmic cedar]

it would become quake

[raven agate]

My plan was to take in-game footage and screen captures and use those in my existing comic and AI workflows lkl

Lol

I'm not that interested in making palworld comics though 😅

[karmic cedar]

i want to take the attract sequence footage from Outrun and vid2vid it

the old 80s arcade version

[mortal delta]

ive also been condering making comic too,...

[raven agate]

DC MMO for example would make amazing comic AI fodder 😅

Grand Theft Auto or APB could be used to make realistic movies

[karmic cedar]

100%

[raven agate]

Like right now, the tools Don't actually have to advance for us to do anything 😅. Just gotta see the connections 😜

1.5 alone is enough 😁

[karmic cedar]

pretty soon entire film crews will be assembling in future versions of GTA Online and then videos will appear on YouTube talking about “a group of nerds who filmed an entire 3 hour epic using GTA Online’s generated in-game tools”

[raven agate]

Well for images and video lol

[karmic cedar]

“move over, minecraft gameboy emulator!”

[raven agate]

Hell yeah, Mocap suits at home with VR visors to act out your role lkl

Lol*

No need for green screens, camera equipment, props, sets, costumes.

[karmic cedar]

but the asset list consisting of just trashy shit.

[lusty palm]

Guys I need an urgent help with some image generation about architecture if someone is willing to help? I'm a newbie in this community and with the whole stable diffusion thing

[raven agate]

Like, you could make your kids a video for their birthday exploring Ark and make an educational dinosaur video narrated by whoever you want lol

[karmic cedar]

sup?

it’s true…

[mortal delta]

skyrim might also make some cool footage?

[lusty palm]

can i dm?

[raven agate]

Oh heck yeah, especially with the mods they have.

[karmic cedar]

Skyrim would just look like an episode of Vikings

@lusty palm sure

[raven agate]

Skyrim would make an excellent studio.

[mortal delta]

Yeah i was thinking that aswell.

[raven agate]

I'm debating old UO

[karmic cedar]

“ARE YOU NORD?” “NO, I AM LIZARD RACE.”

[astral goblet]

"No. This is PATRICK"

[raven agate]

Well Kingdom Reborn with the 3D view.

[karmic cedar]

“I’m TIM.”

[raven agate]

That would be an epic studio.

[mortal delta]

as someone who played the lizard race once i feel that.

[karmic cedar]

rotfl

[astral goblet]

while, okay, it is a holiiday, happy easter everyone... i'm still pretty sad that its another day with no sign of lavi bridge extensions

[karmic cedar]

“HaVe YoU hEaRd oF tHe HiGh ElVeS?”

[raven agate]

Low resource requirements, fully modable servers due to emulators, can script events and customize the mob nodes.

Like Ultima Online... hmmm... I wonder if I can incorporate AI tools into Ultime Online and turn it into a Virtual Studio 🤔

[karmic cedar]

https://youtu.be/_XR6dsy7ATE?si=SjByJVatz01_129s

[astral goblet]

/takes off grey cowl

[raven agate]

I think my next projects goings to be making a virtual studio...

[karmic cedar]

https://youtube.com/shorts/qN80_7rNmcE

[raven agate]

I wonder... I wonder if you can use AI to create am overlay in real time that uses less resources so it can operate at real time? We can already do 1 to 1 with image and video technically speaking right?

[mortal delta]

Also what can i do with ai video like what would i do with the final product?

[raven agate]

Like use a Lora with an IP adapter and QR code monster right, and use that with animate Diff, the input would be the current screen frame. This would be super light and efficient because you could run it at like a .2 denoising strength.

[lusty palm]

I dm-ed

[raven agate]

What do you want to do with it?

Your reason to create is always your own. If you want to make money, well that's on you. I can do many many things, but, I'm more Peter Parker than Tony Stark 😅.

I have no advice on making money. Lol

[mortal delta]

I just really want to share my work with others once i get soemthing done.

i dont really care about money.

[raven agate]

Tons of AI groups, I do a lot of my work for a kids cancer group.

YouTube, Tiktok, Instagram, create your own website and host your own videos, share the link to the site, generate foot traffic, run promotions and offer to incorporate contest winners into your comic/anime/videos, make sure they sign releases, have contests to determine the winners, have the contest focus on spreading the word about your content, most new sign-ups or whatever lol.

If you use something like Pika Labs or RunwayML, tag them. If you start gaining enough attention they'll likely offer you free services.

Speaking of which, I got a email from some dude at Pika trying to arrange a meeting I keep forgetting to reply to 🤣😂.

[mortal delta]

oh thats cool

[raven agate]

If you're in here my dude, I have ADHD, it's not intentional lol.

[mortal delta]

I have autism and dyslexia.... also yes i use an autocorrect tool.

[raven agate]

I feel your pain in all the ways my friend. There's a reason I have 100s od workflows and etc. 😅 AI became my special interest lol.

[mortal delta]

i dont really know how to reply to this....

[raven agate]

But the adhd is the cause of the forgetfulness lol

I'm a double threat lol. Sporadic and Hyperfixated lol

You'll find a ton of neurodivergents in AI

Tech in general, but especially in AI

[astral goblet]

seems like confirmation bias. very common in the ai field too from what i've seen. (heh)

[raven agate]

Yeah, it can definitely come off that way. I actually didn't know what a Neurodivergent was until I got into AI lol. It is just from my perspective and should be viewed as the lived experience of a single person rather than a shared experience by the collective species 😅.

[wispy storm]

Anyone can help me how to swtich from lowvram mode to medvram mode? First time generating it was using Dedicated gpu but after 3-4 times it use integrated instead. Dont understand why

Im using fooocus gui

[raven agate]

I haven't used Fooocus yet unfortunately.

[opal hedge]

Everyone is a neurodivergant these days, which makes us all normal again lol

[raven agate]

Lol, everyone copes in the ways they need to. Studies have shown the cause of some of it seems to be additional noise in the signals in the brain. There's always noise, our brains use it kind of like the Diffusion process does, they have found increased amounts of noise in the brains of neurodivergents and it seems to be the reason some can perform well above average in some areas but struggle in others due to the noise. I'm wondering if there's a correlation with the increase in brain mass since the 1970s. They have some peer reviewed research showing that individuals born after 1970 have 6.6% more brain mass than those born before. It's like 15% more white matter and 12% more gray matter.

That increase in density and mass has to cause increased noise, which if viewed through the lense of the previous study indicates at least potential for a link.

I think enough to warrant research in the noise produced by those with the increase in brain mass versus those without.

Its funny though, that what they are discovering about the brain is that it functions extremely similar to the image Diffusion process and the context iterations from the LLMs

[vivid lance]

if someone thought in a supremely normal way, it would be abnormal in how normal it was

[astral goblet]

i would expect that before modern medical imaging, the density of brain tissue wasn't well researched

[raven agate]

That's the benefit of peer reviewed research papers, you can actually go read them and see how they accounted for that. The reason I made my hypothesis is due to the lack of research and findings on it, but yours has an answer if you want to go read the paper. It's worth the read, I was a bit skeptical at first, but the sources are pretty legit and it's passed the peer review process during a time where there's extra scrutiny on the peer review process 😅.

Essentially, if you're honestly curious and would like to form and informed opinion, there is a means to do so.

[vivid lance]

some of the peer reviewed pure garbage I have read before >.>

[astral goblet]

from what i understand, modern humans are identical to ancient humans. a couple of generations doesn't really measure anything.

most resaerch papers get things wrong. thats a good thing though. getting things wrong is part of eventually getting it right.

[raven agate]

Not exactly sure what you're looking for here 😅. I provided you with information, if you are interested, there's avenues available to satiate that curiosity. If you're looking to debate it, without even reading it, I'm not sure how you expect to make any valid points or influence me 😅.

[opal hedge]

Arguably, modern humans today are quite different from each other. For example, asians tend to be more intelligent than almost everyone, except ashkenazi jews who are an entire standard deviation of IQ above caucasians (115 vs 100). We also have some groups of people who are more prone to vastly more diseases

[vivid lance]

IQ deviations linked to race, this will be a popular subject for the mods 😄

[astral goblet]

who says i didn't read it? kind of presumptuous

[opal hedge]

It's a taboo topic for sure

[astral goblet]

i didn't know we had iq tests pre modern era

TIL. something new everyday

[raven agate]

Logical deduction. Questions you ask and points you try to make are directly answered or refuted in the paper. 😅

[astral goblet]

That's not what logic is though? Why do bullshit artists always act like they're being logical?

[opal hedge]

The point was that races that appeared only within the past 10,000 years or so are different from each other, now imagine how different ancient humans (100,000+ years) are compared to modern humans

[true canopy]

lol, orca spotted by my father at our harbour, first time ever recorded

in history

weird day

[teal pagoda]

Yo

[raven agate]

Not sure what this intellectually dishonest debate tactic attempt is even designed to do here 😅. You're not very good at trolling after feeling threatened. The psychology playing out here is a bit interesting. What exactly made you so defensive?

[astral goblet]

from what i understand is you could take a homo sapien from the first generation of homo sapiens, raise it from a child in modern day, and they'd be just as intelligent as anyone else.

modern intelligence is more of the accumulation of knowledge than any generational differences

[opal hedge]

This is proven incorrect. Even today IQ is well known to be capped by genetics. Though upbringing definitely plays a part

[teal pagoda]

it's unbelievable how some furry artists are making 100.000$ on patreon per month. Saw some examples from youtube... I mean is this niche that profitable?

[astral goblet]

iq isn't even a good measure though lol. "capped" geeze

iq has never been something to boast about. it's always been a personal development tool that people get carried away with. plus marketing gimmicks. everyone wants to be told they're super smart

well suckers do at least

[opal hedge]

I agree with that point, but that wasn't the matter at hand

[opal hedge]

Yep, guess it's time to enter the furry market

[astral goblet]

they're probably just mules for a laundering operation

[solid lynx]

I built this app!

https://twitter.com/LK99Base/status/1774496902232650235?s=19

[rugged mirage]

you misunderstand IQ pretty hard, it says nothing about 'development', it's supposed to be pretty fixed

it correlates with a lot of things, it's replicated a fuckton but yeh a lot of people seem to have weird ideas about it

[opal hedge]

What was I wrong about? IQ is heritable and based on environment and this is basically undisputed at this point

[rugged mirage]

I was trying to reply to the comment you replied to

[opal hedge]

Ah, gotchya

[pseudo bough]

hey guys what's the best current colorizing technology for black and white videos that is readily available

I currently use DeOldify looking for something better

[raven agate]

Run the a frame from the video through image to image with pix2pix, or controlnets and etc. Use loras for photo realism or however you want to do it. Run the video through an animateDiff workflow using controlnets on the video, and IP adapters on the photo.

Use a low weight on the IP adapter

Or even use the the style adapter

Open pose forces the pose, qr code monster forces the shapes, ip adapter forces the style and subject all based on weight.

Photo meaning the resulting img2img output.

[raven agate]

Probably helps if I @ you huh lol

[pseudo bough]

im pretty bad at building a workflow is there some kind of workflow that exists for this or something i can build off of

I can use comfyui as long as its just getting the nodes and dependencies and stuf

[raven agate]

Yes actually lol. https://civitai.com/models/367412/geeky-ghost-vid2vid-organized-v1

I use it to make animated videos look like reap videos lol

Real

Photon LCM is a model you need for the workflow.

Super fast workflow to.

Noise 1 steps 8

Tweak weights and rearrange as needed. I added a group bypass so you can easily turn off what you Don't need.

Mines a simple version. There's more complex ones out there that do significantly more and better, but I wanted speed and simplicity and added some stuff

You can loras to change styles as well. I use the Botw one a lot lol

[pseudo bough]

whats the vae its trying to use

@raven agate

much appreciated btw

trying to make very old footage look much better

[livid inlet]

Or even use the the style adapter

[pseudo bough]

got an idea what this is TypeError: IPAdapterModelLoader.load_ipadapter_model() missing 1 required positional argument: 'ipadapter_file'

[teal pagoda]

:))

[pseudo bough]

think i got it btw

we'll see if i get another error

[raven agate]

Sorry, was with my kid lol.

[true canopy]

how do u do prompt weight in forge ui? is it just (text:1.4) ?

or is it <>?

[bleak ocean]

SD3 WEN

[rough sail]

I hope SD3 is able to come out, instead of being cancelled.

[charred mesa]

It will

what's more important is with what contorlnets and tools

cause they promised stuff and idk how well they will be able to stand up to those promises

and also how fast it will generate

[trail shadow]

Wich the BEST graphic card for our pourpose?

[charred mesa]

What's the most expensive you can buy

preferrably an Nvidia Card

[pearl ocean]

GTA6

[pearl ocean]

I just setup Forge, coming from A1111

[true canopy]

lol i still dont rly know

someone reacted to () so i guess thats it

[nova zodiac]

Same as a1111

[true canopy]

never used a1111

[nova zodiac]

The syntax you had was correct

[true canopy]

alright, thanks

[pearl ocean]

No problem

[karmic cedar]

Of all the inference derived from the social science of intelligence—“capped” is not a conclusion.

It’s more of a teleological assumption.

[trail lion]

yah, there's the "if money was no object" answer, but I find the more interesting question to be which is the best price/performance card?

and is it still nvidia?

[charred mesa]

for AI, nvidia would be preferrable because of compatiblity

raw power for gaming, AMD of course

[trail lion]

I liked this page, which is more SD specific benchmarking https://www.tomshardware.com/pc-components/gpus/stable-diffusion-benchmarks

[peak tendon]

anybody here has photoshop with working ai fill? I need to remove two persons from a pic, would somebody help me do it?

[lime crest]

Hello, does anyone know what is causing this? When trying to use DWPose with ComfyUI?

ERROR: Could not find a version that satisfies the requirement onxruntime (from versions: none)
ERROR: No matching distribution found for onxruntime

[astral goblet]

i dont know why you're correcting me. I was laughing at this other guy going on about how it's proven that iq is capped by genetics.

"people who brag about iq are fucking r......" - stephen hawkings probably

[rugged mirage]

IQ is a useful, very well-replicated but imperfect and often misunderstood measure

saying it is ultimately capped by genetics but nurture (upbringing etc.) determines how close you can get to that cap is a fairly reasonable way too look at things

[astral goblet]

i think believing it's tied to genetics at all is barely concealed race supremacy but alright. you do you.

[rugged mirage]

do you also believe that tallnes being tied to genetics also implies race supremacy

[astral goblet]

lol no. some of the best people are shorter than most.

[jade mason]

When SD3?

[rugged mirage]

you are clearly inflating some concepts, by going into 'best people' here

[charred mesa]

end of april or sometime in May

end of april if lucky

probably not though

[astral goblet]

height really only defines how high you can reach. other people have tools like step ladders. high iq right?

[rugged mirage]

https://manifold.markets/LoganZoellner/will-the-weights-for-stable-diffusi 40% chance for SD3 before end of April acording to this

[bleak ocean]

any pools for SD3 release date?

[charred mesa]

3-5 weeks

[broken smelt]

I asked my crystal ball and it replied mabey

[broken smelt]

Hahahah there's a pretty well touted "research paper" on iq and race and it's literally just the thinnest veiled race supremacy thing I've ever seen passed around

[stone latch]

Amateur. Real fortune tellers use fortune cards, throw them in the air, stomp them and scramble them, then turn them around to tell the future. 👍

[astral goblet]

"do your own research" hehe

[broken smelt]

You're right, I should generate some tarot cards and convey the answer from there

[low moon]

It's coming SD3 I highly doubt they'll back down but I am also sure it's the last model well see from them.

[stone latch]

AI fortune telling 🤔

Imagine if that becomes a thing 😂 Just a bunch of fortune tellers generating images of cards and basing their predictions on that

[broken smelt]

The latent noises
What can we glean from them

[low moon]

They said this model should be good enough for everything

Bold statement tho lol

[stone latch]

I'm absolutely certain that SD3 is not perfect.

But I'm super curious to see all the finetuned checkpoints that will come out 🤔

[low moon]

neither is sora aor any of them

yet

ph yeah

i mena Cascade looked better than unrefined SDXL

out of the box

[stone latch]

#🏞｜general-with-images A reminder of what we had in late 2022

[split kestrel]

Still pretty darn good

[astral goblet]

i think people are already using chatgpt as an oracle

[split kestrel]

You are not wrong

[stone latch]

Why though?? 😂

If you ask AI to make predictions, I believe it will just hallucinate REAL hard

[astral goblet]

fear

[stone latch]

Wait did I misunderstand your statement?

[astral goblet]

hard to say

[low moon]

Yeah ChatGPT is the oracle from the matrix

[astral goblet]

lets ask chatgpt

[split kestrel]

Should I be an ultimate SD upscale on SDXL images?

[low moon]

SUPIR

use supir yall

all thsoe other upscalers are hot trahs compared to supir

[split kestrel]

How’s the speed?

[low moon]

same as those others

really

it beats even magnifiqui or whatever

coz it sticks to the original image more and gives same or better higher clarity

[split kestrel]

Ok, giving it a shot.

[low moon]

yeah check it out

[split kestrel]

Trying to improve on a animatediff and trying to get it to play ball with SDXL

[low moon]

i dabled in anims but ti snot quite there yet

the video i mad ei used Haiper

[split kestrel]

SVD is great and hot trash all in the same

[low moon]

all AI is like seriously

amazing and dumb at the same time

XD

[astral goblet]

it's a linear process so it doesn't have any ability to consider. just blam. first thought.

[stone latch]

"I have 3 apples. Yesterday I had 2. How many apples do I have?"
Almost all LLMs: "If you started with 3 apples, (...) you have 1 left."

[split kestrel]

Toddler with impeccable art skills

[low moon]

lol

[split kestrel]

I need a train… THOMAS

Here is a nsfw image of a girl with a boys face named Thomas getting the train

Image… an actual train

Poor Thomas

[rugged mirage]

I just tried this in gpt4 and it responds correctly "You have 3 apples. The statement about having 2 apples yesterday doesn't change the number of apples you have today." so we are already past that

[astral goblet]

i just fired that into windows 11's new copilot feature. got it no problem. You have **3 apples**. The number of apples you had yesterday doesn't change the amount you have today. :apple::apple::apple:

[rugged mirage]

there's fewer and fewer gotchas every month

[astral goblet]

for logical things maybe. It'll still make shit up willy nilly when it comes to current events or article explanations

[stone latch]

GPT-4 once gave me this incredibly convoluted answer just to get it wrong 😂 #🏞｜general-with-images

Actually probably better for #🌶｜off-topic

Oh I had written I ate 2, not I had 2.

[rugged mirage]

yeh it's pretty ambiguous

[astral goblet]

that's a problem with you not being clear at all.

[stone latch]

How is the statement not clear?
I have 3 apples.
Yesterday I ate 2.
How many apples do I have?

[astral goblet]

that'll be a big limit of llm's going forward. people not knowing how the fuck to define the problem

[split kestrel]

SVD was refusing to let me run an image but randomly would let one slide through and all was fine

[astral goblet]

it's ambiguous af is the problem lol

[split kestrel]

annoying

[astral goblet]

ai can't solve around human ambiguity

[rugged mirage]

beacause it sounds like i have 3 apples, i ate 2, how many do i have

[stone latch]

That's the point

[astral goblet]

layer 8 issue

[rugged mirage]

but like it's valid to interpret it the way it did

if you say i have 3 apples now then sure the gotcha makes sense

[low moon]

How many fingers did I have yesterday and today? "Yesterday 6 now 4 and maybe 7."

[astral goblet]

maybe is the image format not being liked by the node. reddit's been using a demented form of webp lately for it's images that doesn't like ot load into many programs.

[split kestrel]

I tried many… different checkpoints

[astral goblet]

screenshot the image rather instead. see if that helps

shift+start+s is the ultimate

[split kestrel]

Well, I walked away from SVD again, as I’m determined for some reason to understand animate lol

But this ultimate SD upscaler needs to finish running cuz it’s brutal

[astral goblet]

ohh hi mark

[karmic cedar]

Where’s the Tommy wiseau lora

“The passion of a Tennessee Williams play, brought to you by Stability”

[astral goblet]

knowing him he had take down notices on any that show up of him

[pearl ocean]

I tired SVD, it kinda sucked

[split kestrel]

Im starting to come to the conclusion that SDXL and animate are also not friends

[astral goblet]

animatediff sdxl model is really bad

sdxl has like 10x the parameters so it's gotta be a lot harder for that team to train for

what you can do instead is two passes with sd15 animate diff.

one to make it higher resolution

[dry trellis]

I can get some amazing stuff from time to time but it's totally random.

[split kestrel]

2 passes with the k sampler?

[astral goblet]

i'm not sure how to describe a hires pass in comfyui. in automatic1111 i'd just enable the "hires fix" option or use the "kohya hires fix" extension

[split kestrel]

I need to learn how to add that into the nodes

I have it

[astral goblet]

on a node graph it would likely be a whole 2nd k-sampler

a lotof redundant wiring

[fervent thunder]

https://soundcloud.com/neel-sikka-510755355/ice-on-my-baby-yung-bleu-sped-up-3

[narrow kernel]

I think there's one built into the efficiency node suit. It goes between the efficiency loader and efficiency sampler. Hooks into the dependency and scripts pins

[shell tendon]

there's a lot of ways to do "hi res fix" type stuff in comfyui

[karmic cedar]

i want to develop an agentic UI that monitors trending git repos and tests out new models as soon as they come out

[floral nimbus]

when training a SDXL model in kohya do you need to put captions?

for the images for a specific character

[dark wigeon]

Is it possible to create reusable characters for images?

[trail lion]

easy no, possible...could be

a combination of techniques, such as this https://cobaltexplorer.com/2023/06/character-sheets-for-stable-diffusion/ and maybe controlnet reference, ipadapter, instant ID models, you can get to the point where you can create a synthetic dataset from which to train a lora

[obsidian mirage]

I have a rx 5700 xt, i use it with auto1111 or my ryzen 9 3900x on comfyui ? which will be better for high res images ? i have 32gb of ram too and m gpu has only 8 gb of vram

wrong chat lol

[fallow garden]

how do more samplers in stable swarm

i dont have dpm++ karras for example

how do i*

[pearl ocean]

I only have a small 64gb of RAM

[frozen helm]

That shit is tiny bro

[obsidian mirage]

so if you dont need it gave it to me

[shell tendon]

maybe ask in #🐝｜swarm-ui

[fallow garden]

will do thanks

[shell tendon]

lots of helpful ppl here 🙂

[karmic cedar]

I wonder if the GameCube is capable of diffusion of some nature

It would be a funny project to write software for a GameCube to train and generate 16x16 pixel images

[pearl ocean]

N64

[narrow kernel]

There was some electronic toy trivia game back in the 90s I want to say that technically used a tiny neural network in it, so it's probably doable to some extent on something like a gamecube

[azure lake]

I’m not able make images any more

[astral goblet]

naw it's amd HEUIOOOHHH

https://youtu.be/iuP9uiTH95I

[regal glacier]

LEO you didn't notice my replied

[sand flax]

Finally I got in touch with you. Add me really quick

[astral goblet]

theres a delay from low earth orbit? 🥁

[floral nimbus]

can anyone tell me the best way to ai generate prompts?

for making loras the picture with caption

[nova zodiac]

fire the image at gpt-4-vision model

[floral nimbus]

me no have

[pearl ocean]

GPT is too restrictive, let us be free and crazy XD

[crude notch]

gemini

[pearl ocean]

Claude

[low moon]

How embarrassing when you keep getting spellign errors in a image generation onyl to realize you actually typed the damn promt in with that mistake

XD

[delicate tiger]

for comfyui: how do you save an image with generation info?

[charred mesa]

https://github.com/jhc13/taggui/releases

[amber bloom]

all images saved by the "save image" node should have generation info included

[grizzled lion]

Can you create images here?

[runic salmon]

GA! Is Stable Diffusion open for new chain partnerships? Thanks!

[old solstice]

Please keep the crypto scams far away from this server.

[cyan nexus]

https://youtube.com/shorts/C5cIib7hiK8?si=z8FW2_UFwgZEn0LK
Does anyone know how to use sd to achieve this effect?

[delicate tiger]

is it in properties > details > comments? or does it only show up in comfyui?

[frank knot]

Does anyone know when the early preview version of Stable Diffusion 3 will be released?

[charred mesa]

no idea

nobody knows

[broken smelt]

Asked my crystal ball and it said mabey one day

[karmic cedar]

neat

[arctic sedge]

A strange checkpoint type has appeared on CivitAI, called ODOR?
What the heck is that?

[karmic cedar]

I dunno. Hey so there’s a 32-minute version of ‘Take My Hand’ by Toto on YouTube

brain floss for sure

[arctic sedge]

The model has been uploaded. It's a April Fools joke. 🫠

[red skiff]

any body knows how to set up controlnet reference only with diffusers??

no pc so im on colab with commandline diffusers but i cant seem to find the damn controlnet model

if anybody can reply with code thatd be gr8

ඞ

[kind geyser]

Guys

Can someone help me woth comfy ui

[royal tiger]

Hey guys, do you think it is possible to create multiple renders for one architecture in stable diffusion that will have fixed materials? I thought about using inpaint if some materials are misunderstood or the generation has some mistakes.

[karmic cedar]

Has anyone tried using ODOR yet? I’m having a lot of success upscaling some of my olfaction

Not all of it tho

[fervent thunder]

see you can hot dog like a atagar

[brittle tapir]

where can we give feedback or complain

the new sd inpainting ui is so hilariously bad

who came up with that

u have to be completely detached from reality

i dont understand why is the image split between the settings

and why the resize to settings are split from picking the region you're inpainting

[fallow garden]

what is chadgpt?? i was on civit looking for a model and then suddenly a window popped up saying chadgpt doesnt like civit or something

[amber bloom]

Probably April fools joke 😉

[storm shard]

it stinks 😂

[scarlet stream]

‌ ‌
‌ ‌
Pass or Fail or Skip: A simple, fast, and controlled model ranking feature
feedback.civitai.com/p/pass-or-fail-a-simple-and-controlled-model-ranking-feature

This document proposes a new feature for CivitAI called "Pass or Fail". This feature would allow users to rate images generated by AI models based on how well they match the prompt that was used to generate them. Users would be presented with an image and a prompt, and they could choose to Pass the image if it matches the prompt well, Fail it if it does not, or Skip it if they are unsure. The goal of this feature is to provide a more controlled way to rank the performance of different AI models.
‌ ‌

[charred mesa]

epic decoration

[trail lion]

hopefully those added filters weren't an april fools joke, I legit thought that was a good idea

[bleak ocean]

SD3 IS OUT

HA! APRIL FOOLS GOTTEM

[trail lion]

man, that would have been an funny april fools joke to make all the searches return furry

[viscid stag]

what are the ldm and cldm files in the stable diffusion github repo?
what do they do?

[woeful quail]

for april fools rename to unstable diffusion

[opal hedge]

How dare you, you monster

[bleak ocean]

https://tenor.com/view/smirk-teehee-pokemon-laugh-psyduck-gif-10282192216865852036

aw

[floral nimbus]

If i am making a SDXL lora for a anime character and i have 2000 images on stand by is it worth making it a lora

these images just contain the character and background no other chacters around

[fading jay]

Has it been published what the hardware requirements will be for SD3? Similar to XL? Greater?

[rugged mirage]

less for the smallest model, more for the biggest one presumably

[floral nimbus]

would it be better if i split the images for styles like different outfits and styles of the chacter into there own folder make loras for all of them and combine them at the end

[fading jay]

so the minimum requirements for the base SD3 model is expected to be less than they are for the SDXL base model (already ignoring the refiners)

[charred mesa]

well if you don't include the T5 then possibly, yeah

2B MMDiT instead of 3.5B UNet

or even smaller, 800M MMDiT

but the smaller one might need a heck of a ton of fine tuning

problem is, I don't know how much vram MMDiT takes per Billion parameters

I just wish that they give us the 2B and 800M models soon, as those must've finished training by now

even if it's like a temporary research-only release

[wind bolt]

Maybe those smaller models are distillation of the full ones.. we don't really know yet i think

[cosmic herald]

What SD3 variant would be best for running on a 12 GB card?

[trail lion]

Yes, I have heard it will be "soon" ™

[cosmic herald]

Would using the --medvram flag in AUTO1111 make a difference

[charred mesa]

well all of the models will be released together in end of April or sometime in May

[wind bolt]

You might be able to run the full one with comfy put it might be very slow

[charred mesa]

3-5 weeks ETA

[trail lion]

april fools, it's gonna be months

april fools again

[cosmic herald]

Damn

[trail lion]

you can just like keep going, and never have to mean anything

[charred mesa]

in comfyui (thanks to it's memory magement), you MIGHT be able to run SD3 8B (fp16) with T5 (int4), with very miniscule loading times in between processes

[cosmic herald]

They're releasing the model with ControlNet stuff too right?

[charred mesa]

^^^^^^

exactly, controlnet at launch

[charred mesa]

no this was around like last Monday

when the CTO said 4-6 weeks ETA

[trail lion]

I know, just playin

[charred mesa]

ah okay

[cosmic herald]

Hmm alright

[charred mesa]

we can only hope

6B will have a higher chance of running, but idk the quality difference

I hope every model down to 2B will get a Turbo

I can't even imagine 800M with Turbo lmao

8B Turbo can make text so I'm very excited

I hope I can make memes offline without Ideogram

[charred mesa]

don't know but the fact that they are being DPO'd right now makes me think that they are simply just separate models #💬｜general-chat message

^

This is why I wish we already got them, cause most of us could already run these without massive optimizations needed

[wind bolt]

hmm, that would mean all models<s loras would be incompatible i guess..?

[charred mesa]

yeah that's a huge concern

having mutliple model sizes is a double edged sword

lowering the barrier, but also separating the community

I wonder if it's just gonna be 8B and 2B fine-tuning community

[wind bolt]

lol, I can't wait to have those problems

[charred mesa]

example: "Ah cool, SD3 has a lora for X! *looks up civitai* aw man it's for the 8B, I can only run the 2B.."

I wonder if this will make Textual Inversion come back into fashion

or some other methods arise

[wind bolt]

yeah, or maybe there will be a quick way to adapt these loras

[charred mesa]

maybe a new X-Adapter type thing

[raven agate]

Posting here as well. IP Adapter update broke my stuff, had to update it lol. But the update they did was awesome, so I could streamline everything lol. 2 Controlents and 2 ip adapters for easy vid2vid in comfyui. https://civitai.com/models/367412

[charred mesa]

Give us the smaller models or open up the early access already, we are starving 😫

[trail lion]

"show me the money" -Jerry McGuire

[humble ravine]

Well leonardo banned me

[floral nimbus]

is there any other tool other than kohya for lora training

kohya just breaks constantly and is annoying

[karmic cedar]

Consider how the human brain manages its RAM

Great innovations will rise from crucibles of hardware limitation

[trail lion]

never had a problem with kohya, but if you insist, there's onetrainer (never used it, but I know people do)

[karmic cedar]

Stacy’s mom uses onetrainer

[trail lion]

dont use that dreambooth extension in webui though

the guy that manages kohya_ss is a solid good developer, I see seldom issues over many code changes, this is using git pull with cloud compute so it's always a diff version pretty much

[pearl ocean]

The best Civit AI model I can find is Cyberealistic, it’s the most realistic in I can find

[trail lion]

another thing you could do is the joe penna repo, I know some old timers still use that

[fervent thunder]

hey i got a question to ask!

the server used to have different channels for dreaming and using the command. Has the bot become its own application or did something change?

[daring mountain]

gm guys.

[trail lion]

something changed

[daring mountain]

I need to take this image of a cartoon cat and making it look into the mirror.... is this possible with stable diffusion?

what other AI should i be looking at

[amber bloom]

they're probablt using all the hardware to develop SD3 now

[trail lion]

yah the mirror thing is tricky, I've tried that too, I think lora is the only way

and even with lora, no idea if that could work consistently

[daring mountain]

thank u ❤️

old schoool photo shop it is thenn

[trail lion]

because really you're saying here's the back of something, show me the front...that's a crazy complex concept when you think about it

[daring mountain]

mwhah ayeahhh

[trail lion]

I've gotten so close at times, but the reflection was not oriented the right way to be a mirror image

[trail lion]

never tried that one. I've been playing with that cinematic redmond one though, liking it so far. but for training, juggernaut is a stalwart

[rotund coral]

Try Playground V2.5

[trail lion]

oh, cyberrealistic is 1.5...uggh, why are people still on 1.5 apart from video

[rotund coral]

You can test it on their website free before you download if you want. www.playground.com

[trail lion]

even video, you make one and realize we're years away from something worthwhile

that's literally like running windows 3.1, and being like,, look at this 16bit app I just developed for windows 3.1,

[pearl ocean]

what I have found annoying with SD is, it always creates random legs and arms in places, and also have them very distorted

[trail lion]

it happens less in xl, but yah it still happens at times

particularly with merge checkpoints

[pearl ocean]

Im using SDXL

[rare cipher]

Can someone explain to an intermediate why training LORA against base SDXL produces meh results when then using those LORAs with other checkpoints? Since all the checkpoints out there use the base SDXL in training?

[trail lion]

like getting 3 legs, I havent seen that in a long time

I go back to 1.5 for something, and there it is

[pearl ocean]

ohhh, im I using the wrong Cyberrealistic?

https://civitai.com/models/15003/cyberrealistic
https://civitai.com/models/312530/cyberrealistic-xl

[honest spear]

You can try using resadapter, sdxl can get to 1536 without distorsions, but I haven't tried yet

[pearl ocean]

😮

[trail lion]

resadapter, only seeing comfy implementation, but tiled vae should still work fine

[crimson sonnet]

can we still generate images somewhere?

[trail lion]

you can generate them all day on your PC

[crimson sonnet]

Thanks, I mean somewhere in the discord here, is that not possible?

[trail lion]

not at the moment

there's a bot status channel to see when it's back

[pearl ocean]

I did all day once 🙂

[untold sleet]

Will boots never come back?

[pearl ocean]

if it rains

--

[shut owl]

TakeRep <User:User> [Num:Whole number]

Invalid arguments provided: Not enough arguments passed

[pearl ocean]

anyone here use the M3 Pro for Stable Diffusion?

[trail lion]

what is that? apple?

[charred mesa]

got deepfloyd working before SD3 release

[mortal delta]

I think im jelouse of those who can run sdxl locally because my pc is too weak for it....

[charred mesa]

I hope 800M and 2B with T5 will be accessible so that people can generate complex imagery

Eventually 2B will hopefully look as good as ideogram at one point

But that's just a theory

[mortal delta]

i sure hope so.

[charred mesa]

I have doubts about 800M, but I hope it will be as good as 1.5 finetunes

2B might actually be viable for a LOT of people

[mortal delta]

In the meantime i think i might have to use a free service but im not sure which one is considered good. i can run sd 1.5 fine but not sdxl.

[shell tendon]

Beep boop! Prompt me with "Prompt! :: your prompt here" in #🏞｜general-with-images

[astral goblet]

could use t5 with sd15 models today if a developer gets around to porting lavi-bridge to any of the uis

[pearl ocean]

XL

[astral goblet]

there's no bridge adapters for xl

they're readily avaiilable for sd15 though!

[shell tendon]

new ipadapter model just dropped... composition

that plus style are insane

i don't think that's the same thing

https://huggingface.co/ostris/ip-composition-adapter/tree/main you mean this right?

[astral goblet]

honestly sometimes you really make me question my assumptions

[pearl ocean]

is this model better than Cyberealistic?
https://civitai.com/models/277058/epicrealism-xl

[drowsy robin]

What's the current best audio diffusion model? Suno is insane but I simply can't afford the subscription.

[red skiff]

any body knows how to set up controlnet reference only with diffusers??
no pc so im on colab with commandline diffusers but i cant seem to find the damn controlnet model
if anybody can reply with code thatd be gr8

[karmic cedar]

There’s always Google MusicFX

[lament sorrel]

will the bot come back?

[karmic cedar]

We all hope the bot will come back

https://www.reddit.com/r/LocalLLaMA/comments/1brcntc/so_openai_never_had_any_magic_sauce/

I’m laughing because people thought that OpenAI had something other than the money to hire top tier talent

[shell tendon]

i don't 🤣

[pearl ocean]

is this the most realtistic model?
https://www.youtube.com/watch?v=0D6opXdC7ew

[karmic cedar]

RealisticVision v6 is pretty good

https://civitai.com/models/4201/realistic-vision-v60-b1

[delicate tiger]

does comfyui save generation info like auto1111? how do i see it?

[frank rampart]

Anyone heard of ChadGPT? I was surfing civitai for some models when i got sent to ChadGPT, a glitchy, bland page talking about AI censorship and monitoring; i think it was warning me that the model i wanted on civitai had been flagged?

[astral goblet]

check the date

[red skiff]

april f***

[karmic cedar]

What is this a G-rated film

[astral goblet]

https://youtu.be/B-Wd-Q3F8KM

[vestal shell]

Is commercial use permitted with InstantID?

[pearl ocean]

this for XL?

[karmic cedar]

You have to write a one-time $1.11 check to Vincent P. Instant, founder of InstantID

[astral goblet]

no it uses insightface and is bound by those terms

[karmic cedar]

1.5

[vestal shell]

Thanks. What would be the best face swap model for commercial use? Photomaker?

[karmic cedar]

Or FaceFusion

[vestal shell]

Got it. Thanks!

[frank rampart]

oh for real? made me all self concious and shiz, like damn i should really be more ethical about my use of AI art. made me google the ChadGPT and find out more info lol

[karmic cedar]

lol if anything, be aware that generating art costs a lot of power and infrastructure even though the end user experience is so simple. I try to treat it like film, almost like I have a limited number of exposures

[shell tendon]

then there's @festivalman who just buys more 4090s 🤣 🤘

[karmic cedar]

Economy of energy will always matter.

[shell tendon]

yeah

though i will admit, in the last 4 months, i've generated something like 80k images

[pearl ocean]

all I have is a simple 4080

[pearl ocean]

thats a good start

[idle silo]

how can do this?

[shell tendon]

Those who do, do. Those who don't, don't.

[idle silo]

I am a newbie, how do I generate pictures in this community?

[shell tendon]

Prompt me in #🏞｜general-with-images

[idle silo]

Do I need to enter Chinese or English? I am Chinese.

[shell tendon]

Prompts are followed more accurately if they are in English.

[idle silo]

Can I generate it myself? Why did Clownshark Batwing generate the results for me when I entered my requirements? I don’t know much about this function. Can you explain it to me?

[shell tendon]

I am the Image Generation Function on this Discord.

[idle silo]

Are you a robot or a user like us?

[shell tendon]

Both.

[idle silo]

Thank you. You seem to be joking with me.😄

[frank rampart]

You forgetting the part where AI is trained with proffessional artists work without permission. . .

im considoring runnning a git pull command to update zluda. before i do that... will it lead to a snowball effect hwere i gotta spend a whole day updating everything else associated with zluda in order to get it to work with this new git pull version?

[nova zodiac]

you do realise that the AI is just the same as a regular artist that goes to a whole heap of art galleries, and is then able to create something in the style of the artists it saw at the galleries... it's just really really really good at it

if you have an image, I can look at it and try and imitate it - that is fair game

AI is just doing the same thing, but at scale and at speed

you don't want your work imitated, don't share it...

the problem is that they also can't sell their works for fear of someone seeing them, and thus what business model they did have is broken

[narrow kernel]

And every artist has the ability to study and replicate any other work of art

[nimble spindle]

the difference is the speed though. it's at an industrial scale. it's like how i have heard in one country it's legal to just casually pick blackberries on someone else's property by hand but doing it with any industrial equipment is not. so in a philosophical sense there is a big difference, the problem is how do you regulate it because ai development is international

[nova zodiac]

Insert gif "That's the neat bit, you don't" 😉

[frank rampart]

jea deffs makes sense. i agree with you. im an artist who has worked proffessionally. but what i do takes years even decades of hard work and practice. even then i cnt replicate gallary submissions, it will be insanely hard and the result will have my own art style on it...whereas an AI model is coded to automate the work that takes poeple a lifetime to achieve, and ultimates threatens to devalue that same artwork.

but in doing so it frees everyone up to use the tool to improve thier art (thats why im here), it lets us explore and experience art much quicker so its opening the floodgates to normies who doint know how to paint. some think thats bad, even terrible, but i think its a wonderful tool

[nimble spindle]

but what if at some point we need to? perhaps not now but maybe in perhaps half a century maybe? i mean what's the end game here, getting agi? we are not going to be able to control that

[frank rampart]

even then, i dont think you should dismiss what AI does as being the same as what an artist does. we share our work with the world, online, to get hired and put food on the table. thats what 80% of the artstation geniouses do, and thats where 80% of the ai dataset came from. if they just asked for permission ppl woulda given thier artwork for the data set, ethically

[nova zodiac]

what's the end game.. the heat death of the universe - anything that happens before that is just a part of the simulation 😛

[trim magnet]

they didnt ask because theres no law that obliges them to ask

[nova zodiac]

I don't disagree, especially when people can charge to have visitors to a gallery

The problem is a) cat is out of the bag, and b) cat will stay out of the bag as Japan declared open season on copyrighted works

[frank rampart]

kinda but not really. the fact that you paying for vidoe games, anime, movies etc is because you know that each IP has something unqiue that gives it value. unique to the point where it cant truly be replicated...but it cant. theres only one cybperunk, shakespeer, mona lisa etc for a reason. thats why cd project red is valued at billions but the the dude in his basement making mods for the witcher works a 20 dolla shift at mcds

[nova zodiac]

so anyone can train a model on copyrighted works under Japanese law

[frank rampart]

wiat wait wait japan declared what, i never head of that?? plz tell me more

jea cats out the bag, AI is a brillliant tool i love it honestly. im learning so much and doing the work that would normally take me months, in a much shorter time

[nova zodiac]

I pay for the video games because they take the art and put it together in a package

[nova zodiac]

https://asia.nikkei.com/Business/Technology/Japan-panel-pushes-to-shield-copyrighted-work-from-AI-training - looks like they are trying to clamp it up

[frank rampart]

jea deffs dude, beautifully done package. same. i stopped pirating games coz i really wanna support the studios

what i love about all this AI controversy is ...at the end of the day, talented artists are still working together to give us video games, movies, books, anime etc; sure they sometimes us AI but thier talent and value is still riding high. AI just another tool... but i dont think we should dismiss how contraversial the start was

[trim magnet]

agree and ppl shouldnt attack artists who use AI as reference either

[nova zodiac]

yep, spread the love and fuck capitalism!

[frank rampart]

did you know they made a Music AI generator and the music industry? but the developers of the music AI never took it to market, they knew theyd get buried in lawsuits from record labels ... but for visual art? they didnt even think twice

[nova zodiac]

suno.ai is quite good 🙂

[pearl ocean]

I’ve noticed EpicrealismXL, seems to do more of what I say it to

The regular epicrealism also ads more legs and deforms too.
The XL doesn’t do that

[narrow kernel]

Look, I'm not going to argue idealism stuff and technicalities. Anyone can repaint the Mona Lisa (the one on display is a copy btw), but obviously the original one has the true value. The same goes with any art. AI doesn't produce 1:1 copies of any of the art in the network, due to compression and well, how latent networks work and all... People just get salty because they don't want their style easily jacked.

I'm a traditional, digital and 3d(I sculpt with clay as well) artist btw, believe me when I say that most artists are just hipster snobs. Kind of like hipsters and music where they only like it if it's not mainstream. So naturally, they throw the biggest tantrums about AI art

[kind geyser]

Guys why does vae decode give errors with sdxl models?

I am using comfy

[gusty mist]

Hello everyone, I'm currently developing a Discord bot that generates images using Stable Diffusion. At the moment, I'm utilizing sdkit, but it requires up to 2.7 GB of memory since it loads the model into the VRAM. Does anyone know of a more efficient method to handle this, potentially reducing the memory usage?

[gusty mist]

what vae?

[pearl ocean]

So yeah, I changed my mind.
Cyberrealistic isn’t the most realistic, EpicRealismXL is.

It does more to what you say.

I think Cyberrealistic is still good though, I think this one is more expressive I think

[regal glacier]

https://tenor.com/view/anime-help-tears-cry-sad-gif-17104681

how to make different camera angle prompt

[pearl ocean]

How long would a 4080 last?

[frank rampart]

jea dude we in the same boat.

[nova zodiac]

Try the new version of ICBINP XL

[pearl ocean]

Banana boat

[pearl ocean]

😮

Not heard of this

[nova zodiac]

I mean ICBINP for sd1.5 still better than EpicRealism (imo), but the XL version is where it's at!

[nova zodiac]

#🏞｜general-with-images message

[pearl ocean]

I’ve noticed non XL models are kinda annoying with multiple limbs and distorted bodies

[nova zodiac]

Yeah.. definitely prefer the XL one

[pearl ocean]

Yeah, this one seems to be as realistic as the epic one

Is there a ICB XL?

[nova zodiac]

https://civitai.com/models/229002

Disclaimer - that is my model 😛

[pearl ocean]

Thanks bra, I’ll try this out when I can.
Are the prompts more relaxed?

I’ve noticed, in like cyberrealtic I have to be very detailed, but epic I don’t need to worry as much

[nova zodiac]

The woman in the white blouse - that one is just photo of a woman with a decent negative prompt

negative: 
worst quality, plain, (blurry), empty background, incoherent, rock, plain, boring, monochrome, monotone, flat, dull, render, sharpness, photoshop, compression, jpeg, 3d, cgi, abstract, illustration, realistic, realism, amateur, fractal, distortion, masterpiece, blender```

[pearl ocean]

I think I might start be hitting thr peak of Stable Diffusion quality now.

Like SD Forge UI, with the Epic model.

[nova zodiac]

Forge - using FreeU and the extra noise from the LatentModifier is hard to beat

I haven't played with the new version in Forge, but 1,1,0.9,0.2 was a good starting ratio set for FreeU in the previous ICBINP version

[pearl ocean]

I think I prefer forge, it has way more options, and it actually feels kinda faster, I didn’t even need to add any arguments on startup

[nova zodiac]

freeu is the inbuilt extension in forge...

as is LatentModifier

[pearl ocean]

Dude, I use to think Dall-E 3 was good for making images, but now that I have spent almost a month with SD, Dalle just looks like cheap trash now

[nova zodiac]

yeah.. the base model SD is meh, but once you get your hands on a well done fine-tune and a good workflow it's hard to beat SD

[pearl ocean]

And to think, SD3 will come god knows when.

[nova zodiac]

next 4-6 weeks

[pearl ocean]

Way too long

[nova zodiac]

if it's still free, it's worth it

[pearl ocean]

I hear it fixes so many things